DELTA - A DEep learning Language Technology plAtform

What is DELTA?

DELTA is a deep learning based end-to-end natural language and speech processing platform. DELTA aims to provide easy and fast experiences for using, deploying, and developing natural language processing and speech models for both academia and industry use cases. DELTA is mainly implemented using TensorFlow and Python 3.

For details of DELTA, please refer to this paper.

What can DELTA do?

DELTA has been used for developing several state-of-the-art algorithms for publications and delivering real production to serve millions of users. It helps you to train, develop, and deploy NLP and/or speech models, featuring:

• Easy-to-use
  • One command to train NLP and speech models, including:
    • NLP: text classification, named entity recognition, question and answering, text summarization, etc
    • Speech: speech recognition, speaker verification, emotion recognition, etc
  • Use configuration files to easily tune parameters and network structures
• Easy-to-deploy
  • What you see in training is what you get in serving: all data processing and features extraction are integrated into a model graph
  • Uniform I/O interfaces and no changes for new models
• Easy-to-develop
  • Easily build state-of-the-art models using modularized components
  • All modules are reliable and fully-tested

Table of Contents

• Installation
• Quick Start
• Benchmarks
• FAQ
• Contributing
• References
• License
• Acknowledgement

Installation

Quick Installation

We use conda to install required packages. Please install conda if you do not have it in your system.

We provide two options to install DELTA, nlp version or full version. nlp version need minimal requirements and only installs NLP related packages: Note: Users from mainland China may need to set up conda mirror sources, see ./tools/install/install-delta.sh for details.

# Run the installation script for NLP version, with CPU or GPU.
cd tools
./install/install-delta.sh nlp [cpu|gpu]

If you want to use both NLP and speech packages, you can install the full version. The full version needs Kaldi library, which can be pre-installed or installed using our installation script.

cd tools
# If you have installed Kaldi
KALDI=/your/path/to/Kaldi ./install/install-delta.sh full [cpu|gpu]
# If you have not installed Kaldi, use the following command
# ./install/install-delta.sh full [cpu|gpu]

To verify the installation, run:

# Activate conda environment
conda activate delta-py3.6-tf1.14
# Or use the following command if your conda version is < 4.6
# source activate delta-py3.6-tf1.14

# Add DELTA enviornment
source env.sh

# Generate mock data for text classification.
pushd egs/mock_text_cls_data/text_cls/v1
./run.sh
popd

# Train the model
python3 delta/main.py --cmd train_and_eval --config egs/mock_text_cls_data/text_cls/v1/config/han-cls.yml

Manual installation

For advanced installation, full version users, or more details, please refer to manual installation.

Docker install

For Docker users, we provide images with DELTA installed. Please refer to docker installation.

Quick Start

Existing Examples

DELTA organizes many commonly-used tasks as examples in egs directory. Each example is a NLP or speech task using a public dataset. We provide the whole pipeline including data processing, model training, evaluation, and deployment.

You can simply use the run.sh under each directory to prepare the dataset, and then train or evaluate a model. For example, you can use the following command to download the CONLL2003 dataset and train and evaluate a BLSTM-CRF model for NER:

pushd ./egs/conll2003/seq_label/v1/
./run.sh
popd
python3 delta/main.py --cmd train --config egs/conll2003/seq_label/v1/config/seq-label.yml
python3 delta/main.py --cmd eval --config egs/conll2003/seq_label/v1/config/seq-label.yml

Modeling

There are several modes to start a DELTA pipeline:

• train_and_eval
• train
• eval
• infer
• export_model

Before run any command, please make sure you need to source env.sh in the current command prompt or a shell script.

You can use train_and_eval to start the model training and evaluation:

python3 delta/main.py --cmd train_and_eval --config <your configuration file>.yml

This is equivalent to:

python3 delta/main.py --cmd train --config <your configuration file>.yml 
python3 delta/main.py --cmd eval --config <your configuration file>.yml 

For evaluation, you need to prepare a data file with features and labels. If you only want to do inference with feature only, you can use the infer mode:

python3 delta/main.py --cmd infer --config <your configuration file>.yml 

When the training is done, you can export a model checkpoint to SavedModel:

python3 delta/main.py --cmd export_model --config <your configuration file>.yml 

Deployment

For model deployment, we provide many tools in the DELTA-NN package. We organize the model deployment scripts under ./dpl directory.

• Put SavedModel and configure model.yaml into dpl/model.
• Use scripts under dpl/gadapter to convert model to other deployment model.
• All compiled tensorflow libs and delta-nn libs are in dpl/lib.
• Test, benchmark or serve under docker.

Benchmarks

In DELTA, we provide experimental results for each task on public datasets as benchmarks. For each task, we compare our implementation with a similar model chosen from a highly-cited publication. You can reproduce the experimental results using the scripts and configuration in the ./egs directory. For more details, please refer to released models.

NLP tasks

Task	Model	DataSet	Metric	DELTA	Baseline	Baseline reference
Sentence Classification	CNN	TREC	Acc	92.2	91.2	Kim (2014)
Document Classification	HAN	Yahoo Answer	Acc	75.1	75.8	Yang et al. (2016)
Named Entity Recognition	BiLSTM-CRF	CoNLL 2003	F1	84.6	84.7	Huang et al. (2015)
Intent Detection (joint)	BiLSTM-CRF-Attention	ATIS	Acc	97.4	98.2	Liu and Lane (2016)
Slots Filling (joint)	BiLSTM-CRF-Attention	ATIS	F1	95.2	95.9	Liu and Lane (2016)
Natural Language Inference	LSTM	SNLI	Acc	80.7	80.6	Bowman et al. (2016)
Summarization	Seq2seq-LSTM	CNN/Daily Mail	RougeL	27.3	28.1	See et al. (2017)
Pretrain-NER	ELMO	CoNLL 2003	F1	92.2	92.2	Peters et al. (2018)
Pretrain-NER	BERT	CoNLL 2003	F1	94.6	94.9	Devlin et al. (2019)

Speech tasks

Task	Model	DataSet	Metric	DELTA	Baseline	Baseline reference
Speech recognition	CTC	HKUST	CER	36.49	38.67	Miao et al. (2016)
Speaker verfication	TDNN	VoxCeleb	EER	3.028	3.138	Kaldi
Emotion recognition	RNN-mean pool	IEMOCAP	Acc	60.82	56.90	Mirsamadi et al. (2017)

FAQ

See FAQ for more information.

Contributing

Any contribution is welcome. All issues and pull requests are highly appreciated! For more details, please refer to the contribution guide.

References

Please cite this paper when referencing DELTA.

@ARTICLE{delta,
       author = {{Han}, Kun and {Chen}, Junwen and {Zhang}, Hui and {Xu}, Haiyang and
         {Peng}, Yiping and {Wang}, Yun and {Ding}, Ning and {Deng}, Hui and
         {Gao}, Yonghu and {Guo}, Tingwei and {Zhang}, Yi and {He}, Yahao and
         {Ma}, Baochang and {Zhou}, Yulong and {Zhang}, Kangli and {Liu}, Chao and
         {Lyu}, Ying and {Wang}, Chenxi and {Gong}, Cheng and {Wang}, Yunbo and
         {Zou}, Wei and {Song}, Hui and {Li}, Xiangang},
       title = "{DELTA: A DEep learning based Language Technology plAtform}",
       journal = {arXiv e-prints},
       year = "2019",
       url = {https://arxiv.org/abs/1908.01853},
}

License

The DELTA platform is licensed under the terms of the Apache license. See LICENSE for more information.

Acknowledgement

The DELTA platform depends on many open source repos. See References for more information.